Shrub Detection in High-Resolution Imagery: A Comparative Study of Two Deep Learning Approaches
- James, Katherine M F, Bradshaw, Karen L
- Authors: James, Katherine M F , Bradshaw, Karen L
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , book
- Identifier: http://hdl.handle.net/10962/440326 , vital:73766 , ISBN 9783030955021 , https://doi.org/10.1007/978-3-030-95502-1_41
- Description: A common task in high-resolution remotely-sensed aerial imagery is the detection of particular target plant species for various ecological and agricultural applications. Although traditionally object-based image analysis approaches have been the most popular method for this task, deep learning approaches such as image patch-based convolutional neural networks (CNNs) have been seen to outperform these older approaches. To a lesser extent, fully convolutional networks (FCNs) that allow for semantic segmentation of images, have also begun to be used in the broader literature. This study investigates patch-based CNNs and FCN-based segmentation for shrub detection, targeting a particular invasive shrub genus. The results show that while a patch-based CNN demonstrates strong performance on ideal image patches, the FCN outperforms this approach on real-world proposed image patches with a 52% higher object-level precision and comparable recall. This indicates that FCN-based segmentation approaches are a promising alternative to patch-based approaches, with the added advantage of not requiring any hand-tuning of a patch proposal algorithm.
- Full Text:
- Date Issued: 2022
- Authors: James, Katherine M F , Bradshaw, Karen L
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , book
- Identifier: http://hdl.handle.net/10962/440326 , vital:73766 , ISBN 9783030955021 , https://doi.org/10.1007/978-3-030-95502-1_41
- Description: A common task in high-resolution remotely-sensed aerial imagery is the detection of particular target plant species for various ecological and agricultural applications. Although traditionally object-based image analysis approaches have been the most popular method for this task, deep learning approaches such as image patch-based convolutional neural networks (CNNs) have been seen to outperform these older approaches. To a lesser extent, fully convolutional networks (FCNs) that allow for semantic segmentation of images, have also begun to be used in the broader literature. This study investigates patch-based CNNs and FCN-based segmentation for shrub detection, targeting a particular invasive shrub genus. The results show that while a patch-based CNN demonstrates strong performance on ideal image patches, the FCN outperforms this approach on real-world proposed image patches with a 52% higher object-level precision and comparable recall. This indicates that FCN-based segmentation approaches are a promising alternative to patch-based approaches, with the added advantage of not requiring any hand-tuning of a patch proposal algorithm.
- Full Text:
- Date Issued: 2022
Detecting plant species in the field with deep learning and drone technology:
- James, Katherine M F, Bradshaw, Karen L
- Authors: James, Katherine M F , Bradshaw, Karen L
- Date: 2020
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/160445 , vital:40446 , https://0-doi.org.wam.seals.ac.za/10.1111/2041-210X.13473
- Description: Aerial drones are providing a new source of high‐resolution imagery for mapping of plant species of interest, amongst other applications. On‐board detection algorithms could open the door to allow for applications in which drones can intelligently interact with their environment. However, the majority of plant detection studies have focused on detection in post‐flight processed orthomosaics. Greater research into developing detection algorithms robust to real‐world variations in environmental conditions is necessary, such that they are suitable for deployment in the field under variable conditions. We outline the steps necessary to develop such a system, show by example how real‐world considerations can be addressed during model training and briefly illustrate the performance of our best performing model in the field when integrated with an aerial drone.
- Full Text:
- Date Issued: 2020
- Authors: James, Katherine M F , Bradshaw, Karen L
- Date: 2020
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/160445 , vital:40446 , https://0-doi.org.wam.seals.ac.za/10.1111/2041-210X.13473
- Description: Aerial drones are providing a new source of high‐resolution imagery for mapping of plant species of interest, amongst other applications. On‐board detection algorithms could open the door to allow for applications in which drones can intelligently interact with their environment. However, the majority of plant detection studies have focused on detection in post‐flight processed orthomosaics. Greater research into developing detection algorithms robust to real‐world variations in environmental conditions is necessary, such that they are suitable for deployment in the field under variable conditions. We outline the steps necessary to develop such a system, show by example how real‐world considerations can be addressed during model training and briefly illustrate the performance of our best performing model in the field when integrated with an aerial drone.
- Full Text:
- Date Issued: 2020
Segmenting objects with indistinct edges, with application to aerial imagery of vegetation
- James, Katherine M F, Bradshaw, Karen L
- Authors: James, Katherine M F , Bradshaw, Karen L
- Date: 2019
- Subjects: To be catalogued
- Language: English
- Type: text , book
- Identifier: http://hdl.handle.net/10962/460614 , vital:75969 , ISBN 9781450372657 , https://doi.org/10.1145/3351108.3351124
- Description: Image segmentation mask creation relies on objects having distinct edges. While this may be true for the objects seen in many image segmentation challenges, it is less so when approaching tasks such as segmentation of vegetation in aerial imagery. Such datasets contain indistinct edges, or areas of mixed information at edges, which introduces a level of annotator subjectivity at edge pixels. Existing loss functions apply equal learning ability to both these pixels of low and high annotation confidence. In this paper, we propose a weight map based loss function that takes into account low confidence in the annotation at edges of objects by down-weighting the contribution of these pixels to the overall loss. We examine different weight map designs to find the most optimal one when applied to a dataset of aerial imagery of vegetation, with the task of segmenting a particular genus of shrub from other land cover types. When compared to inverse class frequency weighted binary cross-entropy loss, we found that using weight map-based loss produced a better performing model than binary cross-entropy loss, improving F1 score by 4%.
- Full Text:
- Date Issued: 2019
- Authors: James, Katherine M F , Bradshaw, Karen L
- Date: 2019
- Subjects: To be catalogued
- Language: English
- Type: text , book
- Identifier: http://hdl.handle.net/10962/460614 , vital:75969 , ISBN 9781450372657 , https://doi.org/10.1145/3351108.3351124
- Description: Image segmentation mask creation relies on objects having distinct edges. While this may be true for the objects seen in many image segmentation challenges, it is less so when approaching tasks such as segmentation of vegetation in aerial imagery. Such datasets contain indistinct edges, or areas of mixed information at edges, which introduces a level of annotator subjectivity at edge pixels. Existing loss functions apply equal learning ability to both these pixels of low and high annotation confidence. In this paper, we propose a weight map based loss function that takes into account low confidence in the annotation at edges of objects by down-weighting the contribution of these pixels to the overall loss. We examine different weight map designs to find the most optimal one when applied to a dataset of aerial imagery of vegetation, with the task of segmenting a particular genus of shrub from other land cover types. When compared to inverse class frequency weighted binary cross-entropy loss, we found that using weight map-based loss produced a better performing model than binary cross-entropy loss, improving F1 score by 4%.
- Full Text:
- Date Issued: 2019
- «
- ‹
- 1
- ›
- »